When an AI answer matters, I do not tell it to do better.
That feedback is too vague. I give the model a concrete review job: challenge the answer it already gave me.
Give AI Something Specific to Evaluate
My review usually starts with questions like these:
- Is this the best you can do?
- Are you giving me this with 100% confidence?
- Rigorously grade yourself and defend what you just gave me.
I am not treating those words like a magic prompt. The point is to make the model inspect the output already in front of me instead of vaguely asking for another version.
Make It Find the Weak Parts
I want the model to identify what is weak, missing, assumed, unsupported, or poorly matched to the task I gave it.
Then I can ask it to correct the weaknesses that matter.
That is more useful to me than saying “do better” because the second request has an evaluation target.
Confidence Is Not Proof
This method has an important limit.
Asking AI whether it is 100% confident does not prove the answer is accurate. Asking the model to grade itself does not create independent verification.
OpenAI warns that ChatGPT can produce incorrect or misleading output and can sound confident when it is wrong. That is why important information still needs to be checked against reliable sources.
Use the self-review to find reasons to look harder. Do not use it as proof that the answer is safe to trust.
A Practical Review Loop
1. Challenge the answer you already have
Ask whether it is the best answer the model can produce for the task you gave it.
2. Ask for the weaknesses
Make it identify what is missing, unsupported, unclear, or poorly reasoned in the output.
3. Make it grade the work against the original task
The grade matters less than the explanation. Make the model defend why the answer deserves the score it gives itself.
4. Verify what matters outside the model
Check material facts, sources, numbers, and conclusions before they drive an important business decision.
The Business Decision
AI can generate work quickly. It does not remove your responsibility for deciding whether that work is good enough.
The leverage is not endlessly asking for more output. It is having a repeatable acceptance process for deciding what survives review.
Sources
- Jeff Fargo, Fargo Talks, “Do This Before You Trust AI”
- OpenAI Help Center: Does ChatGPT tell the truth?
- OpenAI: Graders
Transcript
Verbatim from the Fargo Walks clip, as recorded.
Don't be a lazy AI prompter. "Do better" is the laziest freaking thing you can tell AI. Challenge it. Ask it, "Is this the best you can do?" Are you giving me this with 100% confidence? Or a really good one is rigorously grade yourself and defend what you just gave me. Whatever task it is, then you're giving AI something to work with. I'm Jeff Fargo, and I could help streamline and scale your business with AI. Go forth.
Jeff Fargo is the founder of Fargo Factor and host of Fargo Talks.
WHEN THE DECISION MATTERS
When AI output starts driving business decisions
The problem is bigger than a better prompt. You need a repeatable way to decide what AI can produce, what deserves another review, and what still requires human verification. That is the kind of decision system Fargo Factor helps you build.
